# Cursor workspace MCP proof Environment: Linux, real T3 Code dev server and web client, Chromium, 1280 x 800 CSS viewport (DPR 2), dark appearance. Fresh isolated T3 homes, no copied database, injected Cursor API credential. The same callable fixture and prompts were used for both commits. 1. Install each checkout with `vp i`. 2. Create an isolated T3 home with Cursor enabled. If the automatic selection is not advertised by the fresh catalog, add it as a custom Cursor selection on both sides. 3. Start each with `T3CODE_DEV_SEED=0 CURSOR_DATA_DIR= vp run dev --home-dir `. 4. Initialize a Git project containing the supplied `server.mjs`, `.cursor/mcp.json` (from `mcp.json`), and `.gitignore` containing `audit.jsonl`. Add the project through `node apps/server/src/bin.ts project add --title "MCP workspace proof" --base-dir `. 5. Pair the web client and select the Cursor provider. The supplied API credential is already authenticated; browser OAuth sign-in is not needed. 6. Create a fresh thread, send TURN 1 from `prompts.txt`, and record until the response settles. The fixture returns distinguishable MCP-only markers. 7. Run `python3 disable-server.py ` without changing mcp.json, then send TURN 2 in the same thread. Continue recording until settled. 8. Blur the input before still captures. Each retained still was checked against an identical screenshot taken at least 500 ms later. Base: both namespaces report discovery errors; no audit file is created. On the second turn, repro still fails and switchable is absent. Head: both tools return their markers. On the second turn, repro returns its marker and switchable is absent. `head-recorded-audit.jsonl` is the excerpt for the recorded two turns: both processes initialize on turn one, and only repro starts and receives a call on turn two. The screenshots and full two-turn recordings are real app captures. Runtime timing is not a performance comparison. Cursor IDE and Windows were not exercised. AI-assisted; verified with the tests and tools listed above.